Mistral: Voxtral Small 24B 2507

Converts audio input (speech) directly into text transcriptions using an integrated audio encoder that processes raw audio waveforms before feeding them into the language model backbone. The model handles variable-length audio sequences and automatically detects language context from acoustic features, enabling accurate transcription across 40+ languages without requiring explicit language specification. Works with streaming and batch audio inputs up to model context limits.

Transcribes audio in a source language and simultaneously translates the transcribed content into a target language (or multiple targets) within a single forward pass. The model uses a shared audio encoder that extracts language-agnostic acoustic features, then routes them through language-specific decoder heads trained on parallel multilingual data. This architecture avoids cascading errors from separate transcription-then-translation pipelines.

Analyzes audio input to extract semantic meaning, intent, emotion, speaker characteristics, and contextual information beyond raw transcription. The model processes audio through its integrated encoder to generate rich embeddings that capture prosody, tone, and acoustic patterns, then applies language understanding layers to infer speaker intent, sentiment, topic, and metadata. Supports queries like 'summarize the key decisions from this meeting' or 'extract action items and assign them to speakers'.

Generates coherent text responses conditioned on audio input, maintaining semantic and contextual information from the audio throughout generation. The model encodes audio into a fixed-size representation that is injected into the language model's hidden states, allowing the decoder to generate text that directly references, summarizes, or responds to audio content. Supports use cases like generating meeting summaries, answering questions about audio content, or creating follow-up messages based on conversation context.

Accepts simultaneous audio and text inputs in a single API request, allowing developers to provide context, instructions, or supplementary information via text while the model processes audio content. The model's architecture supports interleaved audio and text tokens, enabling prompts like 'Transcribe this audio [AUDIO] and answer the question: [TEXT]' or 'Summarize this meeting [AUDIO] focusing on decisions about [TEXT TOPIC]'. Text and audio are encoded through separate pathways and fused in the model's hidden layers.

Processes audio input as a continuous stream rather than requiring complete file uploads, enabling low-latency transcription and analysis of live audio sources (meetings, broadcasts, phone calls). The model uses a streaming encoder that processes audio chunks incrementally and generates partial transcriptions as audio arrives, with optional refinement as more context becomes available. Supports WebSocket or HTTP chunked transfer encoding for continuous audio delivery.